Fujitsu AI Slashes Legacy COBOL Documentation Time by 97%

TL;DR

• Fujitsu launches Application Transform powered by Kozuchi — a SaaS service that analyzes legacy COBOL code and auto-generates design documents with 97% faster turnaround.
• The tool delivers 95% improvement in comprehensiveness and 60% better readability compared to manual documentation, using Knowledge Graph-Enhanced RAG to minimize hallucinations.
• Service goes live in Japan on March 30, 2026, with plans to expand to code rewriting and maintenance support later this fiscal year.
• Targets enterprises stuck maintaining decades-old COBOL systems without adequate documentation — a critical blocker for modernization projects.

Fujitsu Bets on Knowledge Graph-Enhanced RAG for COBOL Analysis

Fujitsu announced the launch of Fujitsu Application Transform powered by Fujitsu Kozuchi, a generative AI service designed to analyze source code from legacy systems and automatically generate design documents. The service goes live in Japan today, March 30, 2026, targeting enterprises wrestling with undocumented COBOL systems that have run for decades.

The tool slashes design document generation time by 97% compared to manual processes. It also delivers a 95% improvement in comprehensiveness and a 60% boost in readability, according to Fujitsu’s internal testing.

The service builds on Fujitsu’s 2025 software analysis offering but adds a proprietary twist — Knowledge Graph-Enhanced RAG specifically tuned for software engineering. That architecture aims to minimize hallucinations when parsing complex legacy code, a problem that plagues general-purpose AI tools when they encounter decades-old COBOL logic.

Why COBOL Documentation Still Matters in 2026

Here’s the thing about legacy systems. They don’t die — they just accumulate more dependencies.

COBOL remains shockingly prevalent in enterprise IT, especially in banking, insurance, and government. These systems process trillions of dollars in transactions annually. But the engineers who wrote them retired years ago, and the documentation — if it ever existed — has long since vanished into filing cabinets or bit-rotted network drives.

That creates a nightmare for modernization projects. You can’t migrate what you don’t understand. And reverse-engineering a million-line COBOL codebase by hand takes months or years, assuming you can even find developers who still read the language fluently.

Fujitsu’s service attacks that bottleneck directly. Feed it your legacy source code, and it generates structured design documents that explain what the system actually does — data flows, business logic, module dependencies, the works. It’s like hiring a team of archaeologists to excavate your own infrastructure, except they work at machine speed.

The 97% time reduction isn’t just a productivity win. It’s the difference between a modernization project that takes three years and one that takes three months. That compression changes the economics of migration entirely.

And the 95% comprehensiveness improvement matters more than it sounds. Incomplete documentation is worse than no documentation — it gives you false confidence. You think you understand the system, migrate half of it, and then discover a critical business rule buried in some obscure subroutine that your incomplete docs missed. Now you’re debugging production at 2 a.m.

Fujitsu’s Knowledge Graph Play vs. General AI Tools

Fujitsu isn’t the first company to point generative AI at legacy code. GitHub Copilot can explain code snippets. ChatGPT can summarize functions if you paste them in. But those tools treat code as text — they don’t model the deep structural relationships that define how a system actually works.

Fujitsu’s Knowledge Graph-Enhanced RAG approach is different. It builds a graph of the codebase — modules, data structures, call chains, dependencies — and uses that graph to ground the AI’s analysis. That structure acts like guardrails, reducing the chance the model hallucinates a plausible-sounding explanation that’s actually wrong.

That distinction matters enormously when you’re analyzing COBOL systems with millions of lines of code and business logic that encodes decades of regulatory changes and edge cases. General AI tools can give you a summary. Fujitsu’s tool aims to give you a map.

The competitive angle here is clear. Fujitsu is betting that enterprises won’t trust general-purpose AI for mission-critical modernization work — they need tools purpose-built for software engineering, with accuracy guarantees and domain-specific architectures. If that bet pays off, it positions Fujitsu as the go-to vendor for AI-assisted legacy migration in Japan and potentially beyond.

Legacy Modernization as a Service Market Heats Up

Fujitsu’s launch signals a broader shift in how enterprises think about technical debt. For years, the standard playbook was to hire consultants, staff up a modernization team, and grind through the migration manually. That approach works, but it’s slow and expensive.

AI-assisted tools flip the equation. If you can auto-generate design docs in days instead of months, suddenly modernization projects that were economically unviable become feasible. That unlocks a massive backlog of deferred migrations — systems that companies knew they should replace but couldn’t justify the cost.

Fujitsu reportedly plans to expand the service to include code rewriting and maintenance support later in fiscal year 2026. That roadmap makes sense. Documentation is step one. But the real prize is automated code translation — taking COBOL and transpiling it to Java or Python with high fidelity. If Fujitsu can nail that, they’re not just selling a productivity tool. They’re selling an escape hatch from decades of technical debt.

The Japan launch is strategic. Japanese enterprises are notoriously conservative about system changes, but they’re also sitting on some of the oldest and most complex legacy systems in the world. If Fujitsu can prove the tool works in that environment, it’s a strong signal for adoption elsewhere.

But the service also faces real challenges. Accuracy is everything. A 95% comprehensiveness improvement sounds great until you hit the 5% of edge cases the tool missed — and one of those edge cases turns out to be a critical tax calculation or fraud detection rule. Enterprises will test this thing ruthlessly before trusting it with production systems.

First question to watch: how does Fujitsu handle liability when the AI-generated docs are wrong? If a migration fails because the tool missed a dependency, who owns that risk — Fujitsu or the customer? That’s not a technical question. It’s a contract negotiation, and it’ll shape adoption.

Second thing to monitor: how fast Fujitsu ships the code rewriting feature. Documentation is useful, but it’s still a manual bottleneck — someone has to read those docs and write new code. Automated translation is where the real 10x productivity gain lives. If Fujitsu can deliver that in 2026, they leapfrog the competition. If it slips to 2027, others catch up.

Third angle: does this stay Japan-only, or does Fujitsu push for global expansion? The COBOL problem is universal. American banks and European insurers face the same documentation nightmare. But enterprise SaaS adoption patterns vary wildly by region, and Fujitsu’s brand strength outside Japan is uneven. Partnerships with global systems integrators could accelerate reach — or Fujitsu could stay focused on dominating the domestic market first.

Source: Fujitsu Press Release